Trust-aware denoising autoencoder with spatial-temporal activity for cross-domain personalized recommendations

Abstract

Recently, cross-domain recommendation systems have been very helpful in improving the quality of recommendation and solving the problem of cold start and data sparsity. Cross-domain recommender systems allow the transfer of knowledge from one domain with dense ratings to other domain with sparse ratings. Such transfer of knowledge helps in addressing the data sparseness and cold start issues in traditional recommender systems. Although cross-domain recommendations have evolved significantly, yet employment of various factors, such as time, trust, and location remains a challenge. Most of the existing approaches ignore the important fact that at what specific time the user may be interested in the recommended item. Moreover, a person’s trust level and sentiments may be influenced by the variation in the location and time, thereby affecting the decision making. In this paper, we propose a cross-domain recommender system that not only takes into account the time at finer granularity levels (e.g., hours, days, weeks, etc.), but also considers a persons location, trust level, and sentiment analysis while computing recommendations. Our proposed model, named as, Trust-Aware Spatial-Temporal Activity based Denoising Autoencoder (TSTDAE), employs autoencoder-based deep-learning models to generate top-N recommendations for a given user and addresses the cold-start problem in the cross-domain scenario of ‘User Overlap’. The proposed work is fivefold: i) Filter out the users’ biased profiles based on sentiment analysis. ii) Learn the features using autoencoder and then perform clustering among the users of source and target domains to discover the best neighbors. iii) Compute the trust and ratio of preference bias between active user (the user to whom top-N items are recommended) and their neighbors and grade the neighbors based on unbiased preferences iv) Project the best time for recommending the items to an active user and generate the top-N recommendations. We have evaluated the proposed model on a public dataset of e-commerce retail service ‘AliExpress’ for the evaluation metrics: Precision, Mean Absolute Precision (MAP), Normalized Discounted Cumulative Gain (NDCG), and Hit Ratio (HR). The experimental results showed improved performance of the proposed system over the existing models.

Introduction

The evolution of Internet technologies has led to the development of a large number of online e-commerce and social-networking applications. In the past couple of years, the volumes of data have increased dramatically due to the onset of COVID-19 pandemic that motivated a large number of people to go for online purchases. This has posed new challenges for information retrieval systems that not only need to search the required content from diversified sources but also the content needs to be most relevant to an individual’s preferences.

Recommender systems have been adopted as viable solutions to address information overload problem as their aim is to suggest only relevant items to a user matching with an individual’s preferences.

Despite significant progress in collaborative filtering-based recommendation systems, there are still open challenges related to data sparseness and cold start that affect the performance of recommender systems. Data sparsity occurs when there are fewer entries provided by users for each item in user-to-item matrix. This results into large number of zero similarity values, leading to irrelevant recommendations for a user. Cold start problem arises when a user is new to the system and the system does not have a sufficient records for computing future recommendations.

Numerous methods have been devised in the past to address the aforementioned challenges. One approach is to embed the trust metrics in recommendation models that helps alleviating the drawbacks of cold start problem. According to sociology, trust is a set of expectations shared by all those involved in the interactions and/or networks [59]. Several trust-aware recommendation systems have been developed recently that generate more specific recommendations for a user depending upon the users’ context such as time, location, and current trust level. A users’ context may be static or dynamic [63]. For instance, trust or location of a user may vary with time. When a user’s context is encoded in the item- and user-specific latent factors, this helps a recommender system to solve the key issues, such as sparsity and cold start [76].

Numerous deep learning-based recommendation architectures have been proposed recently to address the data sparsity and cold start problems in traditional recommender systems [22], [67]. Deep learning effectively captures the non-linear relationships between users and items and provides more complex data abstractions in higher layers [24]. Moreover, from rich sources of data such as textual, and visual information, deep neural networks effectively capture the complex relationships within the data itself [41], [75]. Recommender systems, using cross-domain information, present a new dimension for solving cold start problem by shifting knowledge from source to target domains. One of the advantages of cross-domain systems is that information can be learned from dense domain to sparse domain, on the basis of a similarity score that is computed based on features of items or users [32].

Several works, such as [16], [20], [67], [36] adopted the collaborative filtering approaches for solving sparsity and cold-start problem in recommendation systems. Collaborative filtering (CF) uses similarity values computed based on users' preferences in the form of explicit feedback, usually rated from 1 to 5 scale [65], [31]. However, some users are likely to express their opinion in the form of text. Reviews in the form of text provide better guidelines for any new user to choose or get to know about items’ information.

One of the predominant challenges in e-commerce is the textual feedback biasness that is caused by personal interests of users, current trends, and the choice of users in terms of social influence. Due to this biasness, the performance of recommender systems is affected in terms of accuracy. For instance, Fig. 1, shows the biasness in the reviews of AliExpress¹ given by different users belonging to ‘Russia (RU)’ in the domain of ‘Apparel Accessories’. It can be observed that both users have given 5-star ratings where first user’s review is positive (A***P) and second user’s review is negative (V***r), thus introducing biasness. Sometimes, rating systems are targeted by spammers who falsely modify the ratings of certain products [29].

Most of the existing CF systems ignore the role of users’ textual feedbacks’ despite that these having significant importance in overall product rating, especially when enough numerical ratings are not available [31]. So, it is necessary to bridge the gap between recommendations generated based on ratings and reviews entered as free text, to reach the unbiased rating of items.

In recommender systems, trustworthiness is one of the major issues, for example, malicious activity can change the preferences information that affects the recommendation accuracy [14]. When users want to buy products, then they may have more trust in the feedback provided by their friends (neighbors) as compared to ordinary users. Since the accuracy of rating prediction depends upon the neighbors, therefore, selection of best neighbors who have similar interests in cross-domain, is a challenge [78].

Moreover, recommendation time is another important factor not usually considered by existing literature. It defines the instant when a user may show interest in a particular item. To increase their sales, retailers may want to observe the purchase patterns of certain group of buyers with respect to time. This makes it a challenging problem for most of the existing recommender systems that do not consider the time factor at a finer granularity (e.g., hours, days, weeks, etc.) [68], [36]. By mining a user’s time-variant purchase patterns, we can predict a user's spatial–temporal preference activity. For example, Charlie is interested in buying apparel accessories, so a recommendation of apparel accessories in a nearby POI at a specific time would be convincing.

To address the above-mentioned challenges, our motivation is to provide an effective approach with maximum recommendation accuracy for generating trust-aware recommendations by assuming the scenario of ‘Users Overlap’ in cross-domain e-commerce systems. We have introduced the concept of preference bias which means that users have matching preferences in the regions that are frequently visited by the same users. The assumption is that an improved top-N recommendation can be generated if the users have a strong preference bias in their frequented regions and may have more trust in each other if they belong to the same region.

The following is the summarized list of our contributions:

• We propose a recommendation model called Trust-Aware Spatial-Temporal Activity based Denoising Autoencoder (TSTDAE) that models nonlinear relationships based on autoencoder in cross-domain scenario of ‘User Overlap’ for solving a cold start problem.

• To address feedback bias and improve the accuracy of recommendations, we perform sentiment analysis and filter out the biased users’ preferences based on negative polarity scores.

• For generating features, autoencoder is used by considering the unbiased users’ profiles and then performing clustering for finding neighbors for an active user to whom top-N items are recommended.

• The concept of trust degree and ratio of preference bias is introduced between trustors and trustees and the system ensembles this trust degree and ratio of preference bias with Denoising autoencoder for improving the recommendation quality.

• We project the best time of recommendations for an individual user or group of users using clustering on time series data and compute the preferable time of users’ purchases.

• We evaluate the performance of TSTDAE on public dataset of ‘aliexpress.com’ using Precision, MAP, NDCG, and HR.

The rest of the paper is organized as follows. Section 2 describes the related work. In section 3 we discuss the proposed model. Section 4 explains the experimental setup, Section 5 describes evaluation and results and Section 6 discusses the conclusions and future work.